Dell Technologies World 2026 | Arun Narayanan & Alison Biers, Dell Technologies

Clips
More from Dell Technologies World 2026

Arun Narayanan

SVP, Server and Networking Products

Dell Technologies

Alison Biers

Global Marketing Sr. Director for Server, Networking, & Vertical Industries

Dell Technologies

play_circle_outline 100% Liquid Cooling for Rubin Generation: Advanced Cold-Plate Designs, CDU Innovations, Two-Phase and Immersion Solutions

play_circle_outline Air-cooled server portfolio remains for customers unable to re-plumb their data centers.

play_circle_outline Lightweight SONiC on Tomahawk, Broadcom, NVIDIA Spectrum: Solving Networking Bottlenecks and InfiniBand Needs for Scale Architectures

play_circle_outline Rack-Scale Acceleration: PowerRack, HBM-to-SRAM Memory Tiers, KV Caches, and Fast Storage

play_circle_outline Security: silicon root of trust, iDRAC attestation, and post-quantum cryptography readiness.

Info

Arun Narayanan & Alison Biers, Dell Technologies

Arun Narayanan

SVP, Server and Networking Products Dell Technologies

Alison Biers

Global Marketing Sr. Director for Server, Networking, & Vertical Industries Dell Technologies

search

Dave Vellante

>> Hi everybody. Welcome back to Dell Tech World 2026. I'm Dave Vellante. We're here in Las Vegas at the Venetian and this is day three of our CUBE coverage. John Furrier's here, and Gemma Allen as well joining me is Arun Narayanan, who's the senior vice president of server and networking products at Dell Technologies and Ali Biers, who's the senior director of Global Marketing for Compute Nord Networking and Industries. Dell Technologies. Folks, good to see you again.

Alison Biers

>> Great to see you.

Dave Vellante

>> Thanks so much for spending some time with us. Wow, what a difference a few years make.

Alison Biers

>> Isn't that the truth?

Dave Vellante

>> The portfolio is just rocking. Take us through, Ali, the arc of that portfolio of the last few years. We used to just be here talking about servers. Now we're talking about a much larger unit of performance.

Alison Biers

>> It's so true. Yeah. So if you saw some of the announcements from day one, we talked a little bit about Dell PowerRack. So at the highest echelon of performance and density, we have our new PowerRack portfolio. So not only have we been doing that for compute at rack scale for the last 15 years, but now we're doing it for networking and also for storage. So you're able to pack a lot into those singular racks. And networking, as you know, has really become the next bottleneck. So we're really focused on our networking portfolio. We've been growing that over the last couple of years and it's really starting to grow exponentially because people have such a focus on it.

Dave Vellante

>> I got a tour of one of your AI labs a couple years ago and it was warm in there. Arun, your job is to make sure this stuff doesn't melt in it.

Arun Narayanan

>> That is exactly right.

Dave Vellante

>> It works reliably. So talk a little bit about the engineering work that goes in to developing these systems.

Arun Narayanan

>> Yeah. I'll start with how we've been thinking about portfolio design in general. We begin to understand that the market is changing dynamically very fast. Go back two years ago, the largest, most powerful rack was 80 kilowatts. Come to Vera Rubin, you're going to get racks of 235 kilowatts, and then get to the next generation of Rubin Ultra and Kyber, you're going to very quickly get to one megawatt racks. So you have to fundamentally redesign everything from power distribution to cooling. So at Dell, we've been focused on thinking about how does cooling change? What happens to cooling? So now we've gone to... If our first generation of liquid cooling was like hybrid cooled, 80% of the design was cooled, 20% was not cooled. In the next generation with Vera Rubin and everything we're doing in the portfolio, we are at 100% liquid cooled solution. That means there's nothing in the component that's not touched by cold plate. And we've also innovated on the cooling technology itself, the cold plate design. We've innovated on CDU designs, which is the coolant distribution unit getting into this. What is the pressure that's needed? What is the temperature that's needed? And then we've innovated on... We are working continuously innovating on the materials that are needed for the coolant itself, for the design of the cold plates, for the entire ecosystem end to end. Our teams are looking ahead into the next generation. We're looking at two phase cooling technologies. Do we need those? Are we looking at further refinement in immersion cooling? Do we need those? So our CTO and our research teams are five years ahead right now.

Dave Vellante

>> Okay. That's interesting because when you guys get deep, you even go down to the integrity of the connection of the hose, the tolerances of the hose. You really got to think about those things because if you get a leak in these multi, multi-million dollar systems, what a shame that would be. And now two phase, that's interesting. That's kind of back on the table because that requires different type of fluids and so that requires new engineering thinking. So you guys are ahead of that curve thinking about the next couple of years.

Arun Narayanan

>> I think two phase is three years out, but we have to do the work now to get to those technologies productized three years from now. So our teams are continuously looking at where the next frontier is and pushing towards that.

Dave Vellante

>> Okay. So Arun was just talking about 100% liquid cool. Now not everybody can do the liquid cools. Obviously, the Neoclouds, they're building from a greenfield, Ali, but most enterprises aren't in a position to re-plumb their data centers.

Alison Biers

>> So true. Yeah.

Dave Vellante

>> So they're limited on power. And so how are you guys addressing that?

Alison Biers

>> From an air cool perspective, we still have a full portfolio. And as you just saw, we announced a lot of new air cooled servers, including the R9825, and 15 that'll be coming this year. So we have the AMD EPYC processors in those and generally the portfolio just continues to grow. You'll see a lot around consolidation trying to do a little bit more with fewer servers. You can squeeze more into the core count now and with the servers with that incredible density and performance, it really has advanced so far.

Dave Vellante

>> So networking, you said, has become the new bottleneck. We know that. So we have to keep these systems in balance. Okay. So extreme parallel computing comes up and then the network becomes the bottleneck. NVIDIA's got the NV link 72 scale up, scale out, scale across. InfiniBand, everybody said InfiniBand's dead. It's anything but dead. Arun, take us through the networking technology, how that's advanced, and how you're keeping things in balance.

Arun Narayanan

>> Yeah. Great question. Right. I'll step back and say, first of all, let's start with what you said, right? Network architecture's changed. Go back five years ago. There's one big monolithic network, an enterprise network, data center network. Now with AI, there's three different networks. There's a scale up network that's within the rack. How it scales memory coherency across that is step one. Step two is scale out network, east west network that connects 100,000 GPU cluster to each other. How is that performing? And that network is some very key requirements. It has to be really high bandwidth because you can't have a GPU starving for data. Then you're giving up a lot of infrastructure waiting for data. Second has to be low latency, especially in an inference world where you have all of these agents talking to the other, you need millisecond response. So you need a very fast network. The third question is these training costs are becoming so large that you can't fit in any one data center when you need scale across networking to do that. So our focus is on focusing on scale out networking, focusing on scale up networking, and focusing on scale across networking. And from a technology standpoint, the focus has been on the right operating system. Where do we need an AI operating system to be? And our strong, strong belief is it does not need to be the heavy legacy networking operating systems. We need lightweight, high bandwidth, low latency network operating systems, and we believe SONiC is the future there. SONiC is the Linux of networking. Linux transformed data center computing. I think SONiC will transform AI networking. So we are going to do solutions with SONiC with Tomahawk and Broadcom. We're doing it with SONiC on spectrum, with NVIDIA spectrum network. And of course, there are some customers will still see an InfiniBand, like you said, and will support the InfiniBand infrastructure. So we are doing it all with SONiC operating system.

Dave Vellante

>> Okay. So the extreme, the accelerated so called accelerated computing comes out, it requires new thinking about networking, and then that affects IO. So now we're seeing a new IO. I think of it as a new hierarchy because we got all these levels of memory now. We got high bandwidth memory, we got SRAM, we got low power memory, we got KV cache. So that's like a whole new... The pyramid lives.

Arun Narayanan

>> The whole memory hierarchy is being redefined as we speak.

Dave Vellante

>> Right. So what does that mean for you guys in terms of how you design these PowerRacks?

Arun Narayanan

>> So when we think of PowerRack, to me the PowerRack has to be able to support compute, server, regular server, AI compute, regular compute, storage and networking. And AI system has all of these things. Now if you think of the GPU compute, in gen one, it was all about GPU compute. But when you come to an agentic system, you're going to need CPU compute because agents run on CPUs. They do multiple tool calling. They're going to need CPUs to do that. The second is as this KV cache grows, you're going to need layers and layers of storage. Not the entire KV cache is not going to fit in HBM and the GPU. It's not going to fit in the server memory. You're going to need fast storage to go do that. So we are building CMX and STX with NVIDIA as technologies. In addition to that, all of these agents are going to generate tons and tons of data. That needs to be stored, that needs to be archived. And so we are going to launch object scale storage as part of our exascale framework into the PowerRack. So our PowerRacks will support all of this. And if you're a data center operator, the beauty of this is you can have the same data center topology, be it compute, be it networking, be it storage, be it AI. That could make a data center's life a lot easier. A homogeneous data center with heterogeneous infrastructure. That is what we're going for.

Dave Vellante

>> So Ali, one of the things that is important, you guys stress a lot is customer choice. So we're seeing networking diversity. You got Tomahawk, Ethernet, Spectrum X Ethernet, InfiniBand. Silicon diversity is another customer requirement. So how are you thinking about that? How are you guys positioning that for customers?

Alison Biers

>> Yeah, it's so important to give customers choice. And today we have an incredibly broad portfolio to help address that across the board. I want to go back though to something Arun just said about the rack scale piece because it's not just about having the right technology, it's also about that turnkey ability. Everything has really been engineered together as a system and that's why we're the number one rack scale vendor now this year ranked by IDC. It's why our customers from delivery timeframe to full production in six hours, it's really an incredible feat, and they can choose that level of service with the PowerRack system.

Dave Vellante

>> But this is why I asked you about the diversity, because how do you accommodate that level of diversity for silicon networking, et cetera, and be able to just power it up, spin it up, get into production within six hours. That used to take longer than that to spin up a server, just get it off the dock took six hours.

Arun Narayanan

>> You want me to take that?

Dave Vellante

>> What's the magic sauce?

Arun Narayanan

>> Yeah. So the magic sauce is Dell has been working for the last three and a half years in perfecting what I would call rack scale design, right? And it takes everything from designing the right thing upfront. So we have these FDE engineers who work with our customers to design the solution right. Once we have the design right, what we are doing is we are sending an enormous amount of time in our factories testing all of this, testing each, every component, that's legacy from Dell, but also testing the entire rack with tons of diagnostics and burn-ins. Once we do that, we are able to test... These tested racks are shipped directly to the customer dock and then we are able to just power it on because we have a deployment services arm that is fully trained to take the Dell tested rack and deploy to the customer side. We've had enormous success in deploying this at gigawatt scale data centers across the board with this model. And I think this model is differentiated and this is why we're winning so much in the market today.

Dave Vellante

>> One of the important things is, of course, security. What is this marketing narrative around security? And I want to get into the technical aspects of it.

Alison Biers

>> Of course everybody's talking about quantum and what that means. So in our new generation of PowerEdge, we'll have PQC readiness with everything that we're developing. So customers need to be thinking three to five years out. And so obviously, with our next generation, it's going to announce over the next two years with products time to market, with our latest CPUs, and a lot of other improvements as well to the platform. So I think you just see people really curious around the topic and a lot of fear, but also interest in being strategic about how they plan for the future.

Dave Vellante

>> Well, I think it's important to point out. Shor's algorithm was created, I think in 1994, but quantum computers, once they are able to run Shor's algorithm, we'll be able to break traditional cryptography. So what the adversaries are doing is they're exfiltrating encrypted data with the hopes that, at some point, they'll be able to decrypt.

Alison Biers

>> They'll be able to unlock. Yeah.

Dave Vellante

>> So post quantum cryptography is an important topic to think about now, not wait until quantum computers come out at scale whenever, 2029, 2030, whenever it is. You got to think about that now.

Alison Biers

>> Exactly. And I think everything also that we're doing in the highest part of our business, we try to bring that engineering down into the other products that we have. So as you talk about AI for enterprise as well, a lot of that innovation around power and cooling goes into our products. And I don't know if you notice, but we announced our first CDUs this year that are Dell engineered. It's our IP the first time we've had two product offers in this space. So really exciting.

Dave Vellante

>> Okay. On the security front, people are talking about securing agents, they're talking about securing the data around the agents, but what does it mean from an infrastructure standpoint? How are you rethinking security?

Arun Narayanan

>> Yeah. I think security is going to hit everywhere in the stack, all the way from the top of the agent to the bottom of infrastructure. Let's start with the fact that Dell's infrastructure is going to data centers across the world. Models are going to run on that infrastructure. How does the model trust that this infrastructure is secure, it has not been tampered with, it has attestation services. So we have built our silicon root of trust all the way from buying the silicon, through the factory, to the end model deployment. Second, we built in our iDRAC systems and our systems management ability to attest the service. So you can actually come back and say, "Is this what was purchased? Is this what was sold? Is this what was shipped? Is this what is the exact design that you asked for?" And that the data center operator and the model operator can have high security and attestation back to Dell saying that this is the design that we purchased. We are 100% secure. That is the level of security we are putting in our hardware. In addition to that, we are doing post quantum cryptography across the board, right? So that's what we are designing for.

Dave Vellante

>> We talked a little bit about this, but I want to come back to it, the whole GPU, CPU. The market is so confused. I love to watch the talking heads on TV. It's like, oh, GPU, NVIDIA. Oh, CPU, Intel, AMD. It's so interesting. Let's try to set the table. So we know that the ratio of GPU to CPU has been compressed.

Arun Narayanan

>> Emerging.

Dave Vellante

>> Yeah. So it was eight to one, people say it's four to one, maybe going to two to one. I think Lisa said it could potentially go to one to one, who knows. But all of these companies have CPUs, all your partners do. What's driving that ratio down? What's the CPU built for? What's the GPU built for? How do you get the most utilization out of that GPU, that expensive resource?

Arun Narayanan

>> I think what you're seeing is two macro trends. Trend number one is agents. And trend number two is what I call reinforcement learning. Let's take trend number two first and then we'll come to trend number one second. In reinforcement learning, in generation one, models were getting pre-trained and the model was going out. With reinforcement we're learning, we have this whole sandbox approach where the model throws out there are multiple models running at the same time. You get a sandbox, you get an answer, validate the answer. All of that validation is a CPU workload. So if 70% of all your new models are going to reinforce learn and then the weights of the model get optimized based on that, that is going to drive a whole bunch of new CPU workload. So in training, historically it has been this one to four or one to eight ratio. I think that'll fast converge to one to two to one to one ratio. That's one of the big workload trends that is driving more CPUs. Now on agentic, when you look at it, the agent itself is a body of software that runs on a CPU. It has to orchestrate and run sub-agents. So think of a model where you have an orchestrative agent, we have 15 different agents. All of these 15 different agents are running on CPUs and they're calling multiple tool calls. They're calling an MCP server for something else. They're calling a Gmail server for something else, and then they have to call a model. So all of that is another increased CPU workload. So again, the ratio is going to change and these new agentic architectures are probably going to go from... If you think of gen one of inference, it was us typing on a chatbot and getting answers. Gen two was these reasoning models or thinking models where you would type on a chatbot, then the agent, the model itself would type in a lot of tokens and get responses. So that changed the CPU GPU ratio. Now in gen three, agentic inference, you're going to see multiple agents running on CPU type things and talking to the agent. So that's going to change the CPU, GPU ratio. So I think the trend is converging from one to eight or one to four to one to two to one to one. Who knows? We'll see. But those are the architectures that are evolving.

Dave Vellante

>> Interesting. You think it'll flip to the favor of CPU?

Arun Narayanan

>> I think it will converge. Let's say that in five years, we have 15 million data center GPUs being deployed in a year. I think it's realistic to expect that 15 million CPUs are being deployed against it. And you and I both know this, Dave. If there's one thing in this market, it changes it faster than anything else and what we predict today might change tomorrow.

Dave Vellante

>> But what's interesting is I think people mistakenly thought, oh, CPU's dead. It's all about the GPU. I know when I talk to the major research labs, Livermore, Oak Ridge, et cetera, and they're putting in quantum computers, they'll say, "Oh, we're taking high performance computing AI, which is GPUs, and now quantum." These are hybrid architectures that will live. And we talk about this at events like supercomputing, right? It's not just replacing that. It's .

Arun Narayanan

>> It's one or the other. It's an and. It's not an or in my mind, right? Everything is going to grow as we go forward.

Dave Vellante

>> But the AI factory is the new unit of infrastructure and it's manufacturing intelligence via tokens. We heard Jeff Clark talk about the token is going to be a line item on your P&L. I think he's absolutely right. We're going to be spending more on technology monetizing those tokens. Do you think that the general purpose functions in the install base, which is largely defined by the X86, security, networking, storage, do they get absorbed into the AI factory stack and become one logical unit or do you see them as separate entities?

Arun Narayanan

>> I think it's going to be workload-specific and customer-specific. If you think of a general purpose company, they do some AI workloads and then they have a lot of general purpose workloads. Think of email and file and block. I don't think those things can merge at any point in time in the near future. They're two distinct architectures. But if you're a high performance computing or bond AI native company who's doing AI only type workloads or scientific computing or scientific research, I think there's a much more tighter convergence. So I think it's workload specific. I don't think there's a one aha, this is the future designed for everything. You're going to have-

Dave Vellante

>> Something that we're wrestling with because people talk about deterministic, the SaaS apocalypse, SaaS businesses melting down, but we don't see it that way. We see those worlds working together, the probabilistic and the deterministic and actually you need both.

Arun Narayanan

>> You need both.

Alison Biers

>> You need both. And I think it's a modernization story, because they have to look at their air cooled data centers a little bit differently. Right now they have a lot of servers taking up a lot of space and if they can get five to one, seven to one more than that consolidation in the next generation, it's really going to free up the space so that they can actually do some of those things.

Dave Vellante

>> And do you agree, Ali, that what's new this time around the cloud era... Dell was essentially refreshing, that was the business. It was, okay, we got to refresh cycle better TCO, faster, better, cheaper. Now the modernization is not only the refresh, it's also to support AI workloads. And so there's a new... This transformation of business, there's the tokenomics and you're supporting that. So there's not just a refresh cycle, it's actually supporting revenue generation. It always has, but in a new way.

Alison Biers

>> Yeah. I think people are having trouble sometimes getting their minds around it, too, because in the past it was just straight line depreciation. Every five years, you order your new systems and now what we're finding is when you really dig into it, sometimes it makes sense for them to break their regular depreciation cycle because of the TCO, because the advancements in power and cooling have come so far. And especially from a performance perspective, they can get more out of their systems when they go to the next gen.

Dave Vellante

>> Well, it's funny, when Michael Burry came out with his short NVIDIA and it's the end of these things have such a short life, I wrote a piece and I said, "You rethink the useful life." Maybe it's going to compress a little bit. I think I was wrong. I think today's training infrastructure becomes tomorrow's inference infrastructure. You can actually see elongated depreciation cycles. I know all the cloud guys have elongated their depreciation cycles. They may not have to shrink them.

Arun Narayanan

>> I think the hardest asset to find out there's an H100 GPU. You can't find one.

Dave Vellante

>> Yeah. Right. Try getting it on eBay. Good luck.

Arun Narayanan

>> It's hard to find it. Even all H100 capacity at every new cloud is already sold out. The point is the life cycle of these assets are much, much longer than we ever thought it would be.

Dave Vellante

>> And I think what people are realizing is you can do a lot.

Arun Narayanan

>> On an H100 .

Dave Vellante

>> N minus two generation GPUs and the CPUs now that are coming along with them. Ali, did you get all the product stuff in? Did I miss anything? .

Alison Biers

>> I think so. I think one more thing to note just on the networking front. Now what's unique is we're actually selling the switches from NVIDIA directly through Dell. So for our customers, I think it's really important because now they have that single point of contact for support and we do that model really well for customers.

Dave Vellante

>> Well, Jensen has made comments on the earnings call. We're the biggest networking company by revenue, which is-

Arun Narayanan

>> I don't think he's wrong about it. I think he's right.

Dave Vellante

>> I was like, wow, I'm looking at Cisco and I'm like, wow, it's pretty close.

Arun Narayanan

>> Data center networking. I think NVIDIA is the biggest company on there.

Dave Vellante

>> I think so. But well guys, congratulations. It's very exciting times. Dell is just rocking. The infrastructure is just amazing. Hardware matters still more than ever. So thank you so much.

Arun Narayanan

>> I think hardware is more sexy now.

Dave Vellante

>> It totally is. I love it. Thanks so much, guys.

Alison Biers

>> Great to be with you. Thank you.

Dave Vellante

>> All right. Keep it right there. Thank you for watching. This is Dave Vellante for John Furrier and Gemma. Gemma Allen, we're here at Dell Tech World 2026. We're live. Day three, you're watching TheCUBE.

Arun Narayanan & Alison Biers, Dell Technologies

search

Dave Vellante

>> Hi everybody. Welcome back to Dell Tech World 2026. I'm Dave Vellante. We're here in Las Vegas at the Venetian and this is day three of our CUBE coverage. John Furrier's here, and Gemma Allen as well joining me is Arun Narayanan, who's the senior vice president of server and networking products at Dell Technologies and Ali Biers, who's the senior director of Global Marketing for Compute Nord Networking and Industries. Dell Technologies. Folks, good to see you again.

Alison Biers

>> Great to see you.

Dave Vellante

>> Thanks so much for spending some time with us. Wow, what a difference a few years make.

Alison Biers

>> Isn't that the truth?

Dave Vellante

>> The portfolio is just rocking. Take us through, Ali, the arc of that portfolio of the last few years. We used to just be here talking about servers. Now we're talking about a much larger unit of performance.

Alison Biers

>> It's so true. Yeah. So if you saw some of the announcements from day one, we talked a little bit about Dell PowerRack. So at the highest echelon of performance and density, we have our new PowerRack portfolio. So not only have we been doing that for compute at rack scale for the last 15 years, but now we're doing it for networking and also for storage. So you're able to pack a lot into those singular racks. And networking, as you know, has really become the next bottleneck. So we're really focused on our networking portfolio. We've been growing that over the last couple of years and it's really starting to grow exponentially because people have such a focus on it.

Dave Vellante

>> I got a tour of one of your AI labs a couple years ago and it was warm in there. Arun, your job is to make sure this stuff doesn't melt in it.

Arun Narayanan

>> That is exactly right.

Dave Vellante

>> It works reliably. So talk a little bit about the engineering work that goes in to developing these systems.

Arun Narayanan

>> Yeah. I'll start with how we've been thinking about portfolio design in general. We begin to understand that the market is changing dynamically very fast. Go back two years ago, the largest, most powerful rack was 80 kilowatts. Come to Vera Rubin, you're going to get racks of 235 kilowatts, and then get to the next generation of Rubin Ultra and Kyber, you're going to very quickly get to one megawatt racks. So you have to fundamentally redesign everything from power distribution to cooling. So at Dell, we've been focused on thinking about how does cooling change? What happens to cooling? So now we've gone to... If our first generation of liquid cooling was like hybrid cooled, 80% of the design was cooled, 20% was not cooled. In the next generation with Vera Rubin and everything we're doing in the portfolio, we are at 100% liquid cooled solution. That means there's nothing in the component that's not touched by cold plate. And we've also innovated on the cooling technology itself, the cold plate design. We've innovated on CDU designs, which is the coolant distribution unit getting into this. What is the pressure that's needed? What is the temperature that's needed? And then we've innovated on... We are working continuously innovating on the materials that are needed for the coolant itself, for the design of the cold plates, for the entire ecosystem end to end. Our teams are looking ahead into the next generation. We're looking at two phase cooling technologies. Do we need those? Are we looking at further refinement in immersion cooling? Do we need those? So our CTO and our research teams are five years ahead right now.

Dave Vellante

>> Okay. That's interesting because when you guys get deep, you even go down to the integrity of the connection of the hose, the tolerances of the hose. You really got to think about those things because if you get a leak in these multi, multi-million dollar systems, what a shame that would be. And now two phase, that's interesting. That's kind of back on the table because that requires different type of fluids and so that requires new engineering thinking. So you guys are ahead of that curve thinking about the next couple of years.

Arun Narayanan

>> I think two phase is three years out, but we have to do the work now to get to those technologies productized three years from now. So our teams are continuously looking at where the next frontier is and pushing towards that.

Dave Vellante

>> Okay. So Arun was just talking about 100% liquid cool. Now not everybody can do the liquid cools. Obviously, the Neoclouds, they're building from a greenfield, Ali, but most enterprises aren't in a position to re-plumb their data centers.

Alison Biers

>> So true. Yeah.

Dave Vellante

>> So they're limited on power. And so how are you guys addressing that?

Alison Biers

>> From an air cool perspective, we still have a full portfolio. And as you just saw, we announced a lot of new air cooled servers, including the R9825, and 15 that'll be coming this year. So we have the AMD EPYC processors in those and generally the portfolio just continues to grow. You'll see a lot around consolidation trying to do a little bit more with fewer servers. You can squeeze more into the core count now and with the servers with that incredible density and performance, it really has advanced so far.

Dave Vellante

>> So networking, you said, has become the new bottleneck. We know that. So we have to keep these systems in balance. Okay. So extreme parallel computing comes up and then the network becomes the bottleneck. NVIDIA's got the NV link 72 scale up, scale out, scale across. InfiniBand, everybody said InfiniBand's dead. It's anything but dead. Arun, take us through the networking technology, how that's advanced, and how you're keeping things in balance.

Arun Narayanan

>> Yeah. Great question. Right. I'll step back and say, first of all, let's start with what you said, right? Network architecture's changed. Go back five years ago. There's one big monolithic network, an enterprise network, data center network. Now with AI, there's three different networks. There's a scale up network that's within the rack. How it scales memory coherency across that is step one. Step two is scale out network, east west network that connects 100,000 GPU cluster to each other. How is that performing? And that network is some very key requirements. It has to be really high bandwidth because you can't have a GPU starving for data. Then you're giving up a lot of infrastructure waiting for data. Second has to be low latency, especially in an inference world where you have all of these agents talking to the other, you need millisecond response. So you need a very fast network. The third question is these training costs are becoming so large that you can't fit in any one data center when you need scale across networking to do that. So our focus is on focusing on scale out networking, focusing on scale up networking, and focusing on scale across networking. And from a technology standpoint, the focus has been on the right operating system. Where do we need an AI operating system to be? And our strong, strong belief is it does not need to be the heavy legacy networking operating systems. We need lightweight, high bandwidth, low latency network operating systems, and we believe SONiC is the future there. SONiC is the Linux of networking. Linux transformed data center computing. I think SONiC will transform AI networking. So we are going to do solutions with SONiC with Tomahawk and Broadcom. We're doing it with SONiC on spectrum, with NVIDIA spectrum network. And of course, there are some customers will still see an InfiniBand, like you said, and will support the InfiniBand infrastructure. So we are doing it all with SONiC operating system.

Dave Vellante

>> Okay. So the extreme, the accelerated so called accelerated computing comes out, it requires new thinking about networking, and then that affects IO. So now we're seeing a new IO. I think of it as a new hierarchy because we got all these levels of memory now. We got high bandwidth memory, we got SRAM, we got low power memory, we got KV cache. So that's like a whole new... The pyramid lives.

Arun Narayanan

>> The whole memory hierarchy is being redefined as we speak.

Dave Vellante

>> Right. So what does that mean for you guys in terms of how you design these PowerRacks?

Arun Narayanan

>> So when we think of PowerRack, to me the PowerRack has to be able to support compute, server, regular server, AI compute, regular compute, storage and networking. And AI system has all of these things. Now if you think of the GPU compute, in gen one, it was all about GPU compute. But when you come to an agentic system, you're going to need CPU compute because agents run on CPUs. They do multiple tool calling. They're going to need CPUs to do that. The second is as this KV cache grows, you're going to need layers and layers of storage. Not the entire KV cache is not going to fit in HBM and the GPU. It's not going to fit in the server memory. You're going to need fast storage to go do that. So we are building CMX and STX with NVIDIA as technologies. In addition to that, all of these agents are going to generate tons and tons of data. That needs to be stored, that needs to be archived. And so we are going to launch object scale storage as part of our exascale framework into the PowerRack. So our PowerRacks will support all of this. And if you're a data center operator, the beauty of this is you can have the same data center topology, be it compute, be it networking, be it storage, be it AI. That could make a data center's life a lot easier. A homogeneous data center with heterogeneous infrastructure. That is what we're going for.

Dave Vellante

>> So Ali, one of the things that is important, you guys stress a lot is customer choice. So we're seeing networking diversity. You got Tomahawk, Ethernet, Spectrum X Ethernet, InfiniBand. Silicon diversity is another customer requirement. So how are you thinking about that? How are you guys positioning that for customers?

Alison Biers

>> Yeah, it's so important to give customers choice. And today we have an incredibly broad portfolio to help address that across the board. I want to go back though to something Arun just said about the rack scale piece because it's not just about having the right technology, it's also about that turnkey ability. Everything has really been engineered together as a system and that's why we're the number one rack scale vendor now this year ranked by IDC. It's why our customers from delivery timeframe to full production in six hours, it's really an incredible feat, and they can choose that level of service with the PowerRack system.

Dave Vellante

>> But this is why I asked you about the diversity, because how do you accommodate that level of diversity for silicon networking, et cetera, and be able to just power it up, spin it up, get into production within six hours. That used to take longer than that to spin up a server, just get it off the dock took six hours.

Arun Narayanan

>> You want me to take that?

Dave Vellante

>> What's the magic sauce?

Arun Narayanan

>> Yeah. So the magic sauce is Dell has been working for the last three and a half years in perfecting what I would call rack scale design, right? And it takes everything from designing the right thing upfront. So we have these FDE engineers who work with our customers to design the solution right. Once we have the design right, what we are doing is we are sending an enormous amount of time in our factories testing all of this, testing each, every component, that's legacy from Dell, but also testing the entire rack with tons of diagnostics and burn-ins. Once we do that, we are able to test... These tested racks are shipped directly to the customer dock and then we are able to just power it on because we have a deployment services arm that is fully trained to take the Dell tested rack and deploy to the customer side. We've had enormous success in deploying this at gigawatt scale data centers across the board with this model. And I think this model is differentiated and this is why we're winning so much in the market today.

Dave Vellante

>> One of the important things is, of course, security. What is this marketing narrative around security? And I want to get into the technical aspects of it.

Alison Biers

>> Of course everybody's talking about quantum and what that means. So in our new generation of PowerEdge, we'll have PQC readiness with everything that we're developing. So customers need to be thinking three to five years out. And so obviously, with our next generation, it's going to announce over the next two years with products time to market, with our latest CPUs, and a lot of other improvements as well to the platform. So I think you just see people really curious around the topic and a lot of fear, but also interest in being strategic about how they plan for the future.

Dave Vellante

>> Well, I think it's important to point out. Shor's algorithm was created, I think in 1994, but quantum computers, once they are able to run Shor's algorithm, we'll be able to break traditional cryptography. So what the adversaries are doing is they're exfiltrating encrypted data with the hopes that, at some point, they'll be able to decrypt.

Alison Biers

>> They'll be able to unlock. Yeah.

Dave Vellante

>> So post quantum cryptography is an important topic to think about now, not wait until quantum computers come out at scale whenever, 2029, 2030, whenever it is. You got to think about that now.

Alison Biers

>> Exactly. And I think everything also that we're doing in the highest part of our business, we try to bring that engineering down into the other products that we have. So as you talk about AI for enterprise as well, a lot of that innovation around power and cooling goes into our products. And I don't know if you notice, but we announced our first CDUs this year that are Dell engineered. It's our IP the first time we've had two product offers in this space. So really exciting.

Dave Vellante

>> Okay. On the security front, people are talking about securing agents, they're talking about securing the data around the agents, but what does it mean from an infrastructure standpoint? How are you rethinking security?

Arun Narayanan

>> Yeah. I think security is going to hit everywhere in the stack, all the way from the top of the agent to the bottom of infrastructure. Let's start with the fact that Dell's infrastructure is going to data centers across the world. Models are going to run on that infrastructure. How does the model trust that this infrastructure is secure, it has not been tampered with, it has attestation services. So we have built our silicon root of trust all the way from buying the silicon, through the factory, to the end model deployment. Second, we built in our iDRAC systems and our systems management ability to attest the service. So you can actually come back and say, "Is this what was purchased? Is this what was sold? Is this what was shipped? Is this what is the exact design that you asked for?" And that the data center operator and the model operator can have high security and attestation back to Dell saying that this is the design that we purchased. We are 100% secure. That is the level of security we are putting in our hardware. In addition to that, we are doing post quantum cryptography across the board, right? So that's what we are designing for.

Dave Vellante

>> We talked a little bit about this, but I want to come back to it, the whole GPU, CPU. The market is so confused. I love to watch the talking heads on TV. It's like, oh, GPU, NVIDIA. Oh, CPU, Intel, AMD. It's so interesting. Let's try to set the table. So we know that the ratio of GPU to CPU has been compressed.

Arun Narayanan

>> Emerging.

Dave Vellante

>> Yeah. So it was eight to one, people say it's four to one, maybe going to two to one. I think Lisa said it could potentially go to one to one, who knows. But all of these companies have CPUs, all your partners do. What's driving that ratio down? What's the CPU built for? What's the GPU built for? How do you get the most utilization out of that GPU, that expensive resource?

Arun Narayanan

>> I think what you're seeing is two macro trends. Trend number one is agents. And trend number two is what I call reinforcement learning. Let's take trend number two first and then we'll come to trend number one second. In reinforcement learning, in generation one, models were getting pre-trained and the model was going out. With reinforcement we're learning, we have this whole sandbox approach where the model throws out there are multiple models running at the same time. You get a sandbox, you get an answer, validate the answer. All of that validation is a CPU workload. So if 70% of all your new models are going to reinforce learn and then the weights of the model get optimized based on that, that is going to drive a whole bunch of new CPU workload. So in training, historically it has been this one to four or one to eight ratio. I think that'll fast converge to one to two to one to one ratio. That's one of the big workload trends that is driving more CPUs. Now on agentic, when you look at it, the agent itself is a body of software that runs on a CPU. It has to orchestrate and run sub-agents. So think of a model where you have an orchestrative agent, we have 15 different agents. All of these 15 different agents are running on CPUs and they're calling multiple tool calls. They're calling an MCP server for something else. They're calling a Gmail server for something else, and then they have to call a model. So all of that is another increased CPU workload. So again, the ratio is going to change and these new agentic architectures are probably going to go from... If you think of gen one of inference, it was us typing on a chatbot and getting answers. Gen two was these reasoning models or thinking models where you would type on a chatbot, then the agent, the model itself would type in a lot of tokens and get responses. So that changed the CPU GPU ratio. Now in gen three, agentic inference, you're going to see multiple agents running on CPU type things and talking to the agent. So that's going to change the CPU, GPU ratio. So I think the trend is converging from one to eight or one to four to one to two to one to one. Who knows? We'll see. But those are the architectures that are evolving.

Dave Vellante

>> Interesting. You think it'll flip to the favor of CPU?

Arun Narayanan

>> I think it will converge. Let's say that in five years, we have 15 million data center GPUs being deployed in a year. I think it's realistic to expect that 15 million CPUs are being deployed against it. And you and I both know this, Dave. If there's one thing in this market, it changes it faster than anything else and what we predict today might change tomorrow.

Dave Vellante

>> But what's interesting is I think people mistakenly thought, oh, CPU's dead. It's all about the GPU. I know when I talk to the major research labs, Livermore, Oak Ridge, et cetera, and they're putting in quantum computers, they'll say, "Oh, we're taking high performance computing AI, which is GPUs, and now quantum." These are hybrid architectures that will live. And we talk about this at events like supercomputing, right? It's not just replacing that. It's .

Arun Narayanan

>> It's one or the other. It's an and. It's not an or in my mind, right? Everything is going to grow as we go forward.

Dave Vellante

>> But the AI factory is the new unit of infrastructure and it's manufacturing intelligence via tokens. We heard Jeff Clark talk about the token is going to be a line item on your P&L. I think he's absolutely right. We're going to be spending more on technology monetizing those tokens. Do you think that the general purpose functions in the install base, which is largely defined by the X86, security, networking, storage, do they get absorbed into the AI factory stack and become one logical unit or do you see them as separate entities?

Arun Narayanan

>> I think it's going to be workload-specific and customer-specific. If you think of a general purpose company, they do some AI workloads and then they have a lot of general purpose workloads. Think of email and file and block. I don't think those things can merge at any point in time in the near future. They're two distinct architectures. But if you're a high performance computing or bond AI native company who's doing AI only type workloads or scientific computing or scientific research, I think there's a much more tighter convergence. So I think it's workload specific. I don't think there's a one aha, this is the future designed for everything. You're going to have-

Dave Vellante

>> Something that we're wrestling with because people talk about deterministic, the SaaS apocalypse, SaaS businesses melting down, but we don't see it that way. We see those worlds working together, the probabilistic and the deterministic and actually you need both.

Arun Narayanan

>> You need both.

Alison Biers

>> You need both. And I think it's a modernization story, because they have to look at their air cooled data centers a little bit differently. Right now they have a lot of servers taking up a lot of space and if they can get five to one, seven to one more than that consolidation in the next generation, it's really going to free up the space so that they can actually do some of those things.

Dave Vellante

>> And do you agree, Ali, that what's new this time around the cloud era... Dell was essentially refreshing, that was the business. It was, okay, we got to refresh cycle better TCO, faster, better, cheaper. Now the modernization is not only the refresh, it's also to support AI workloads. And so there's a new... This transformation of business, there's the tokenomics and you're supporting that. So there's not just a refresh cycle, it's actually supporting revenue generation. It always has, but in a new way.

Alison Biers

>> Yeah. I think people are having trouble sometimes getting their minds around it, too, because in the past it was just straight line depreciation. Every five years, you order your new systems and now what we're finding is when you really dig into it, sometimes it makes sense for them to break their regular depreciation cycle because of the TCO, because the advancements in power and cooling have come so far. And especially from a performance perspective, they can get more out of their systems when they go to the next gen.

Dave Vellante

>> Well, it's funny, when Michael Burry came out with his short NVIDIA and it's the end of these things have such a short life, I wrote a piece and I said, "You rethink the useful life." Maybe it's going to compress a little bit. I think I was wrong. I think today's training infrastructure becomes tomorrow's inference infrastructure. You can actually see elongated depreciation cycles. I know all the cloud guys have elongated their depreciation cycles. They may not have to shrink them.

Arun Narayanan

>> I think the hardest asset to find out there's an H100 GPU. You can't find one.

Dave Vellante

>> Yeah. Right. Try getting it on eBay. Good luck.

Arun Narayanan

>> It's hard to find it. Even all H100 capacity at every new cloud is already sold out. The point is the life cycle of these assets are much, much longer than we ever thought it would be.

Dave Vellante

>> And I think what people are realizing is you can do a lot.

Arun Narayanan

>> On an H100 .

Dave Vellante

>> N minus two generation GPUs and the CPUs now that are coming along with them. Ali, did you get all the product stuff in? Did I miss anything? .

Alison Biers

>> I think so. I think one more thing to note just on the networking front. Now what's unique is we're actually selling the switches from NVIDIA directly through Dell. So for our customers, I think it's really important because now they have that single point of contact for support and we do that model really well for customers.

Dave Vellante

>> Well, Jensen has made comments on the earnings call. We're the biggest networking company by revenue, which is-

Arun Narayanan

>> I don't think he's wrong about it. I think he's right.

Dave Vellante

>> I was like, wow, I'm looking at Cisco and I'm like, wow, it's pretty close.

Arun Narayanan

>> Data center networking. I think NVIDIA is the biggest company on there.

Dave Vellante

>> I think so. But well guys, congratulations. It's very exciting times. Dell is just rocking. The infrastructure is just amazing. Hardware matters still more than ever. So thank you so much.

Arun Narayanan

>> I think hardware is more sexy now.

Dave Vellante

>> It totally is. I love it. Thanks so much, guys.

Alison Biers

>> Great to be with you. Thank you.

Dave Vellante

>> All right. Keep it right there. Thank you for watching. This is Dave Vellante for John Furrier and Gemma. Gemma Allen, we're here at Dell Tech World 2026. We're live. Day three, you're watching TheCUBE.